R0058/2026-04-03/C001 — Assessment¶
BLUF¶
The claim is partially correct. The 83% homophily figure is directly confirmed by Roytburg & Miller (2025), who report 83.1% global homophily across 6,442 AI research papers. However, the "only 1% of authors bridging the divide" mischaracterizes the source: the paper reports that the top 1% of authors by network degree control 58% of cross-disciplinary paths, which is a finding about concentration of bridging influence, not about the total number of bridging authors.
Probability¶
Rating: Likely (55-80%)
Confidence in assessment: High
Confidence rationale: The primary source was fetched and its specific claims verified against the assertion. The 83.1% figure is unambiguous. The mischaracterization of the "1% bridging" finding is clear when comparing the claim's wording against the source text. Three independent sources provide corroborating context. Confidence is high because the assessment rests on direct source verification rather than interpretation.
Reasoning Chain¶
-
FACT: Roytburg & Miller (2025) report 83.1% global homophily in AI safety-ethics co-authorship networks, based on 6,442 papers across 12 major ML/NLP conferences (2020-2025). Statistical significance confirmed via null model comparisons (p<0.001). [SRC01-E01, Medium-High reliability, High relevance]
-
FACT: The claim states "83% homophily," which is an accurate rounding of the measured 83.1%. This component of the claim is confirmed. [SRC01-E01, Medium-High reliability, High relevance]
-
FACT: Roytburg & Miller report that the top 1% of authors by degree control 58.0% of cross-disciplinary paths, and the top 5% control 88.1%. This is a statement about the concentration of bridging influence among high-degree nodes. [SRC01-E02, Medium-High reliability, High relevance]
-
JUDGMENT: The claim states "only 1% of authors bridging the divide," which implies that 99% of authors have zero cross-field activity. The source's actual finding is that a small elite dominates the bridging that exists — not that only 1% participate in it at all. The paper separately reports that mixed papers represent 9.5% of the corpus, further contradicting a "1%" bridging rate. [SRC01-E02, SRC01-E03] [JUDGMENT]
-
FACT: Independent bibliometric analysis finds 94% of AI ethics institutions in the largest connected component, suggesting the broader ecosystem is connected even while specific subfield boundaries are sharp. [SRC02-E01, Medium reliability, Medium relevance]
-
REPORTED: A scoping review of AIES & FAccT articles finds that safety is underrepresented in ethics-focused venues, consistent with community separation. [SRC03-E01, Medium reliability, Medium relevance]
-
REPORTED: A NeurIPS 2019 workshop was dedicated to bridging "epistemic fractures" between fairness, ethics, and safety research, confirming the divide was recognized at major venues years before quantitative measurement. [SRC04-E01, Medium reliability, Medium relevance]
-
JUDGMENT: The claim is partially correct. The 83% homophily component is confirmed. The 1% bridging component is a mischaracterization that materially changes the meaning of the source finding. The overall probability is rated "Likely" rather than "Very likely" because the second component introduces meaningful inaccuracy. [JUDGMENT]
Evidence Base Summary¶
| Source | Description | Reliability | Relevance | Key Finding |
|---|---|---|---|---|
| SRC01 | Roytburg & Miller (2025) — primary study | Medium-High | High | 83.1% homophily confirmed; top 1% by degree control 58% of paths; mixed papers at 9.5% |
| SRC02 | Qiu, Cheng & Huang (2025) — independent bibliometrics | Medium | Medium | 94% of AI ethics institutions in largest connected component |
| SRC03 | Mehrotra et al. (2025) — AIES/FAccT scoping review | Medium | Medium | Safety underrepresented in ethics-focused venues |
| SRC04 | NeurIPS 2019 "Minding the Gap" workshop | Medium | Medium | Epistemic fractures between fairness, ethics, and safety recognized since 2019 |
Collection Synthesis¶
| Dimension | Assessment |
|---|---|
| Evidence quality | Medium — one primary quantitative study (preprint, conference-accepted) with three corroborating sources of lower direct relevance |
| Source agreement | High — all sources agree the safety-ethics divide exists; none dispute the 83.1% figure |
| Source independence | Medium — SRC02, SRC03, and SRC04 are independent of the primary source, but none directly replicate the homophily measurement |
| Outliers | None — no source contradicts the core finding of community separation |
Detail¶
The evidence collection presents a consistent picture: AI safety and ethics research communities are structurally separated, with high in-group collaboration rates. The primary source (Roytburg & Miller) provides the only direct quantitative measurement of this separation. Three independent sources corroborate the phenomenon from different angles: institutional network analysis (SRC02), topic modeling (SRC03), and institutional recognition of the problem (SRC04).
The main limitation is single-study dependence for the specific numerical claims. No replication or contradicting quantitative study was found. This could mean: (a) the finding is so new that replication has not yet occurred, (b) the finding is robust and no contradicting evidence exists, or (c) the evidence base is simply too small to draw strong conclusions. Given the methodological rigor of the primary study (null models, significance testing, large dataset), interpretation (a) is most likely.
Gaps¶
| Missing Evidence | Impact on Assessment |
|---|---|
| No replication of 83.1% homophily measurement by independent researchers | Moderate — a single unreplicated study is the sole basis for the specific number. The methodology appears sound but independent confirmation would strengthen the finding. |
| No data on what percentage of authors actually bridge (as opposed to how much the top 1% dominate) | Low — the absence of a direct "percentage of bridging authors" makes it impossible to determine the correct figure to replace the claim's "1%." The 9.5% mixed-paper rate provides a proxy but measures papers, not authors. |
| Study limited to 12 ML/NLP venues; broader venue coverage could change the homophily measurement | Low-moderate — researchers publishing in interdisciplinary journals or non-ML venues are not captured, which could systematically undercount cross-field work. |
Researcher Bias Check¶
Declared biases: The researcher has a pro-infrastructure mindset that favors structural explanations. A claim about structural community divides aligns with this bias, potentially making the researcher predisposed to accept it uncritically.
Influence assessment: The finding that the claim is partially incorrect — specifically that the "1% bridging" is a mischaracterization — demonstrates that the analysis was not influenced by confirmation bias. The researcher provided the primary source as candidate evidence, which could have anchored the analysis toward full confirmation, but the mischaracterization was identified and documented.
Cross-References¶
| Entity | ID | File |
|---|---|---|
| Hypotheses | H1, H2, H3 | hypotheses/ |
| Sources | SRC01, SRC02, SRC03, SRC04 | sources/ |
| ACH Matrix | — | ach-matrix.md |
| Self-Audit | — | self-audit.md |